GPU-based acceleration of an RNA tertiary structure prediction algorithm

Experimental techniques such as X-ray crystallography and nuclear magnetic resonance have been useful for the accurate determination of RNA tertiary structures. However, high-throughput structure determination using such methods often becomes difficult, due to the need for a large quantity of pure samples. Computational techniques for the prediction of RNA tertiary structures are thus becoming increasingly popular. Most of the existing prediction algorithms are computationally intensive, and there is a clear need for acceleration. In this paper, we propose a parallelization methodology for the fragment assembly of RNA (FARNA) algorithm, one of the most effective methods for computational prediction of RNA tertiary structure. The proposed parallelization scheme exploits multi-core CPUs and GPUs in harmony to maximize their utilization. We tested our approach with a number of RNA sequences and confirmed that it allows the time required for structure prediction to be significantly reduced. With respect to the baseline architecture equipped with a single CPU core, we achieved a speedup of up to approximately 24×(roughly 4× by multi-core CPUs and 20× by GPUs). Compared with a quad-core CPU setup, the proposed approach delivers an additional 12× speedup by utilizing GPU devices. Given that most PCs these days have a multi-core CPU and a GPU card, our methodology will be very helpful for accelerating algorithms in a cost-effective manner.